Pretrained NASNetMobile incorrect inference with torch backend

[fchollet]

All Keras Applications work correctly with the torch backend except NASNetMobile. It runs, but its predictions are incorrect. To investigate.

[chenmoneygithub]

Looked into the issue, seems the forward pass produces numerical diff: https://colab.research.google.com/gist/chenmoneygithub/5d8b8f508763f603c5d39df2f52b0084/test-keras-core-nasnet.ipynb

[chenmoneygithub]

The problem is Nasnet uses average pooling + same padding, which is not well supported in torch backend.

[fchollet]

With a random input in your Colab the numerical diff looks fairly small, but in practice with real-world images it invalidates the predicted classes entirely, which means we won't be able to make the model available for torch.